ResearchPublicationsOpen SourceAboutUpdatesExperienceProjectsKnowledge BaseRésuméContact

Research

Research focus

Retrieve. Execute. Recover.

Agents often look capable on the happy path. I study what happens off it: skill libraries that grow into the thousands, long-horizon tool use, partial failures, and tight context budgets. Each direction pairs a paper with an open implementation.

Research signals

56.7%

fewer tokens with Graph-of-SkillsSkillsBench · GPT-5.2 Codex · vs. full skill loading

6

papers in 2026, two at EMNLP

1.1K+

GitHub stars on Dr. ClawOpen-source AI scientist workspace I help maintain

01Retrieve

Graph-of-Skills

Retrieves a bounded bundle of skills with their prerequisites from an executable skill graph, so agents use less context without dropping what a task depends on.

View the publication: Graph-of-Skills

02Execute & recover

COMFYCLAW

Typed workflow edits, rollback, and visual verification let an image-generation agent detect, diagnose, and repair its own failures, then keep the fixes as reusable skills.

View the publication: COMFYCLAW

03Build in the open

Dr. Claw

Wraps existing coding agents in an auditable, human-in-the-loop workspace, so planning, execution, and writing form one traceable, recoverable research loop.

View the publication: Dr. Claw

Publications

Ordered by relevance to my current research rather than by date; the first three are the projects above. Every card has a one-click BibTeX entry.

* Equal contribution (co-first author). † Core contributor, as designated in the cited paper.

Graph-of-Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills

EMNLP 2026 Main ConferenceACM CAIS 2026 Workshop AgentSkillsApr 2026

Dawei Liu†, Zongxia Li†, Hongyang Du, Xiyang Wu, Shihang Gui, Yongbei Kuang, Lichao Sun

Dependency-aware retrieval keeps prerequisite skills in a bounded context bundle: +7.0 reward points and 56.7% fewer tokens on SkillsBench with GPT-5.2 Codex, versus full skill loading.

Modern LLM agents increasingly rely on reusable skills, and as they interact with personal applications, web browsers, and other interfaces, skill libraries can scale to thousands of skills. Scaling to larger skill sets introduces two key challenges. First, loading the full skill set saturates the context window, driving up token costs, hallucination, and latency. Second, semantic retrieval surfaces topically relevant skills but misses their prerequisite chain of upstream and downstream skills, creating a prerequisite gap that leaves the retrieved bundle execution-incomplete. In this paper, we present Graph-of-Skills (GoS), an inference-time structural retrieval layer for large skill libraries. GoS constructs an executable skill graph offline from skill packages, then at inference time retrieves a bounded, dependency-aware skill bundle through hybrid semantic-lexical seeding, reverse-aware Personalized PageRank, and context-budgeted hydration. On SkillsBench and ALFWorld, GoS consistently delivers substantial reward improvements and token savings across three model families (Claude Sonnet 4.5, MiniMax M2.7, and GPT-5.2 Codex). On SkillsBench with GPT-5.2 Codex, GoS raises average reward by 7.0 absolute points over full skill loading, a 25.6% relative gain, while cutting total tokens by 56.7%. Ablations confirm this pattern across skill libraries from 200 to 2,000 skills.

PaperGitHub

COMFYCLAW: Self-Evolving Skill Harnesses for Image Generation Workflows

ECCV 2026 Workshop MUCGJul 2026

Zongxia Li*, Dawei Liu*, Fuxiao Liu, Yuhang Zhou, Xiyang Wu, Jingxi Chen, Jing Xie, Xiaomin Wu, Lichao Sun

A ComfyUI agent harness that validates graph edits, repairs visual failures, and evolves reusable skills; best average score across all six evaluated agent–backbone settings.

Agents are increasingly used to control executable workflows rather than only answer questions. This makes reliability depend on both execution-time scaffolding and the ability to reuse experience from past runs. We study this problem in workflow-based image generation and introduce COMFYCLAW, an agentic harness that controls an unmodified ComfyUI runtime. COMFYCLAW represents workflow construction as typed graph editing, gates tools by construction stage, rolls back invalid edits, and uses a region-level vision-language model (VLM) verifier to turn visual failures into repair suggestions. It also evolves a progressively disclosed skill library, where trajectories, execution errors, and verifier feedback are distilled into reusable Agent Skills after held-out validation. Across four benchmark splits, three agent models, and two image backbones, COMFYCLAW achieves the best average score in all six agent–backbone settings, improving the strongest setting from 61.09 to 76.34 over a verifier-only baseline without skill evolution. Human annotations further show that annotators prefer COMFYCLAW over variants without skill evolution.

PaperGitHub

Dr. Claw: An AI Scientist Workspace for Vibe Research

EMNLP 2026 System DemonstrationsSep 2026

Dingjie Song, Hanrong Zhang, Dawei Liu, Yixin Liu, Zongxia Li, Zhengqing Yuan, Siqi Zhang, Henry Peng Zou, Zhiling Yan, Yuxuan Zhang, Yanfang Ye, Philip S. Yu, Lichao Sun

A human-in-the-loop AI scientist workspace that makes planning, execution, and writing traceable and recoverable across coding-agent executors.

Command-line coding agents (e.g., Claude Code, Gemini CLI) can already read and write files and sustain long sessions, yet end-to-end research still fragments across chat tools, IDEs, terminals, and writing environments, and the decisions that make it auditable are rarely preserved. We present Dr. Claw, an open-source workspace that wraps existing coding-agent executors in a controllable and auditable human-in-the-loop workflow rather than introducing another autonomous agent. Persistent state objects, a reusable skill library, and multi-executor coordination link human decisions to AI execution, turning planning, execution, and writing into one traceable, recoverable loop. We demonstrate Dr. Claw through an interactive three-view scenario and a failure-recovery walkthrough, and evaluate it against a bare command-line agent sharing the same backend executor, so the comparison contrasts the whole orchestration layer (task graph, state objects, and skill library) with the agent it wraps. Holding the executor fixed, Dr. Claw scores higher on research completeness while persisting an auditable, recoverable process trail. Demo access: repository https://github.com/OpenLAIR/dr-claw, released under AGPL-3.0 with GPL-3.0 upstream components.

PaperGitHub

A Cookbook of 3D Vision: Data, Learning Paradigms, and Application

CVPR 2026 Workshop OpenSUN3D · pp. 8376–8390Mar 2026

Hongyang Du*, Zongxia Li*, Dawei Liu*, Runhao Li*, Haoyuan Song, Qingyu Zhang, Yubo Wang, Jingcheng Ni, Shihang Gui, Congchao Dong, Tao Hu

A data-centric map of 3D vision connecting geometric representations, datasets, learning paradigms, and applications.

3D vision has rapidly evolved, driven by increasingly diverse data representations, learning paradigms, and modeling strategies. Yet the field remains fragmented across representations and benchmarks, making it difficult to develop unified perspectives on efficiency, fidelity, and scalability. This work provides a data-centric taxonomy of 3D vision that connects geometric representations, datasets, learning frameworks, and applications within a single conceptual map. We survey principal structural representations of 3D data, then examine how dataset design, benchmark construction, and supervision regimes shape recent advances spanning 2D-supervised 3D learning, implicit neural representations, and 4D world modeling.

PaperGitHub

Multimodal Video Generation Models with Audio: Present and Future

ACM MM 2026 Workshop JAV-CGMar 2026

Dawei Liu, Hongyang Du, Zongxia Li, Xiyang Wu, Lantao Yu, Jingxi Chen, Fuxiao Liu, Xiaomin Wu, Jing Xie, Chengsong Huang, Yicheng He, Guangyao Shi

A survey of joint video and audio generation, organized around cross-modal fusion, post-training, evaluation, and synchronization.

Video generation has advanced rapidly, yet visual content alone is insufficient for many realistic and engaging media experiences: sound is central to immersion, semantic coherence, and perceived causality. Growing demand for synchronized dialogue, effects, music, and ambience has therefore accelerated research on multimodal video generation, which coordinates video and audio generation or uses one modality to control the other instead of treating sound as a separate post-production stage. Recent proprietary and open-weight systems, including Veo 3.1, Gemini Omni Flash, Sora 2, Seedance 2.0, Seedance 2.5, Kling 3.0, MiniMax M2.6, Vidu Q3, Wan 2.6, OVI, and LTX-2, illustrate the shift toward native audiovisual capabilities. This paper surveys the foundations and recent advances of multimodal video generation. We organize the literature around four themes: architectural components and cross-modal fusion; post-training and evaluation methods; applications and active research areas; and persistent challenges in synchronization, efficiency, data, safety, and long-horizon consistency. We identify multi-timescale alignment, from local event timing to global narrative and acoustic continuity, as a unifying challenge across current systems. This perspective connects progress in media creation with emerging uses in world modeling, interactive simulation, embodied intelligence, and human–AI communication.

PaperGitHub

FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders

arXiv preprint · cs.CVSep 2026

Hongyang Du, Yunfei Xie, Junjie Ye, Jiawei Yang, Xiaoyan Cong, Haodong Zhang, Yongchao Huang, Haiyu Wu, Zongxia Li, Shihang Gui, Dawei Liu, Runhao Li, Jingcheng Ni, Chen Wei, Randall Balestriero, Yue Wang

Training on random subsets of encoder layers lets one representation-autoencoder decoder reconstruct from any layer fusion; swapping in that decoder alone cuts unguided gFID by 27% on ImageNet-256.

Representation autoencoders (RAEs) reuse features from a pretrained visual encoder as reconstruction and diffusion latents, integrating strong visual representations into image generation. However, RAEs still need to decide which encoder layers form the shared latent space for the generator and pixel decoder. This choice involves a trade-off: shallower layers tend to preserve fine pixel details better, while deeper layers tend to yield better generation metrics. A fixed heuristic layer fusion therefore couples two stages that benefit from different information. We introduce FuseReg, which replaces heuristic feature selection with training over random subsets of encoder layers. We theoretically analyze the underlying mechanism: subset sampling preserves the full-layer latent mean in expectation while explicitly penalizing sensitivity to cross-layer disagreement. On ImageNet-256 with DINOv3-L, a single FuseReg decoder reconstructs from full, sparse, and single-layer fusions without retraining, achieving higher PSNR than decoders specialized to fixed fusions. This flexibility also benefits generation: decoder replacement alone reduces unguided gFID by 27% with an unchanged RAEv2 DiT-XL generator. Applying FuseReg to both stages also reduces unguided gFID by 29% on DiT-Base. The reconstruction and generation benefits also extend to other encoder families. FuseReg narrows the reconstruction-generation gap without additional training cost or architectural changes.

PaperGitHubWebsite

TIMEDB: tumor immune micro-environment cell composition database with automatic analysis and interactive visualization

Nucleic Acids Research · 51(D1): D1417–D1424Jan 2023

Xueying Wang*, Lingxi Chen*, Wei Liu*, Yuanzheng Zhang, Dawei Liu, Chenxin Zhou, Shuai Shi, Jiajie Dong, Zhengtao Lai, Bingran Zhao, Wenjingyu Zhang, Haoyue Cheng, Shuaicheng Li

My role: built the visualization framework and several of the interactive analyses.

A database of tumor immune cell composition with automatic analysis and interactive visualization across 39,706 samples and 43 cancer types.

TIMEDB is an online database for analyzing human tumor immune microenvironment cell-type composition from bulk expression profiles. It curates expression and composition profiles with clinical information for 39,706 samples from 546 datasets across 43 cancer types, and provides automatic analysis with interactive visualization.

PaperGitHubWebsite

The complete, up-to-date list is on Google Scholar and DBLP.

Open Source

Repositories that others can run and build on. Each card states my role, from first-author implementation to core maintenance.

Graph-of-Skills

First author · Maintainer2026

The open-source reference implementation of the EMNLP 2026 paper. I proposed the dependency-graph formulation, built the retrieval system (hybrid semantic–lexical seeding, reverse-aware Personalized PageRank, and context-budgeted hydration), and led the evaluation.

PaperGitHub

Dr. Claw

Core contributor · Maintainer2026

An open-source AI scientist workspace (EMNLP 2026 System Demonstrations). As a core contributor and active maintainer, I built the cross-platform Electron desktop app and its release pipeline, and implemented provider integrations, project and session workflows, and reliability fixes.

PaperGitHub

COMFYCLAW

Co-first author2026

A typed-graph harness for image-generation workflows, with validation, rollback, repair, visual verification, and skill evolution built into execution. I built the ComfyUI plugin, the core agent loop for typed edits to workflow graphs, and the experiment framework; I also ran two of the model evaluations and wrote the paper's first draft.

PaperGitHub

Long-Horizon Terminal-Bench (LHTB)

Contributor2026

A 46-task benchmark of whether LLM agents can sustain useful work in a containerized terminal over hundreds of steps, scored with dense rewards by hidden verifiers that rebuild results from artifacts. Tasks span games and puzzles, multimodal analysis, software and reverse engineering, scientific computing, security and performance, and research reproduction.

PaperGitHubWebsite

About

I'm an AI systems researcher and software engineer at ByteDance (Commerce Ads, Seattle), and I collaborate on agent research with Prof. Lichao Sun and OpenLAIR. My work has moved from scientific visualization and distributed tracing to recommendation, agent infrastructure, and evaluation. The constant has been making complex systems useful, inspectable, and dependable.

How I work

I'm drawn to the unglamorous middle between prototype and production.

That middle is evaluation, interfaces, observability, latency budgets, failure handling, and the many small choices that determine whether software holds together. Industry work taught me to care about them; research lets me study them directly.

AI systemsAgents & retrievalExecution reliabilityEvaluation & infrastructure

Research impact & service

Citations

90+ · h-index 4Google Scholar, Oct 2026

Research use

14 papers use, reimplement, compare against, or test components of Graph-of-SkillsSep 2026

Service

Reviewer, ICLR 2027, ACM TIST, and ACM CAIS 2026 Workshop AgentSkillsICLR: five assigned submissions, Oct 2026

Updates

Selected milestones

Newest first

1. Oct 2026
   ServiceReviewing five submissions for ICLR 2027 and a manuscript for ACM Transactions on Intelligent Systems and Technology (TIST).
2. Sep 2026
   PreprintThe Dr. Claw system paper, accepted to EMNLP 2026 System Demonstrations, is now on arXiv.
3. Aug 2026
   PaperGraph-of-Skills accepted to the EMNLP 2026 Main Conference, after an earlier poster at the ACM CAIS 2026 Workshop AgentSkills.
4. Jul 2026
   PaperCOMFYCLAW accepted to the ECCV 2026 Workshop MUCG (arXiv:2607.01709).
5. Jul 2026
   BenchmarkContributed to Long-Horizon Terminal-Bench, a 46-task benchmark for long-running terminal agents released with the LAIR community.
6. Jun 2026
   CareerJoined ByteDance full-time as a Software Engineer on Commerce Ads in Seattle.

Education

Graduate

University of Pennsylvania

Aug 2024 – May 2026Philadelphia, PA

M.S.E.Master of Science in Engineering/Computer and Information Science

GPA3.90 / 4.00

• Hagan International Scholarship

Undergraduate

Northeastern University

Sep 2020 – Jun 2024Shenyang, China

B.E.Bachelor of Engineering/Software Engineering

GPA3.95 / 4.00first six semestersRank 5 / 396 (Sep 2023)

• National ScholarshipMinistry of Education, 2020–2021
• Merit-based ScholarshipAwarded annually
• Outstanding Graduate
• Outstanding Student×3
• Outstanding Student Leader

Experience

Across four engineering roles, I have worked on advertising systems, agent infrastructure, observability, platform engineering, and backend reliability. That experience is why my research questions usually arrive with interfaces, latency budgets, and failure modes attached.

For more on the production agents in my ByteDance role, read Agent Systems in Production.

ByteDance Inc.

Software Engineer, ByteDance Inc.

Commerce Ads

May 2025 – PresentSeattle, WA

Production systems for Shop Ads and AI-assisted engineering, including an evidence-grounded on-call agent, an interoperability layer for coding agents, and a resumable framework for controlled service validation.

Themes

Production agentsAgent interoperabilityValidation autonomyReliability engineering

Software EngineerJun 2026 – Present

• Architected and built a production on-call agent that triages alerts using SOPs, runtime telemetry, change history, and service-topology knowledge, pairing deterministic tools with a bounded LLM behind guardrails and an evaluation framework.
• Built a multi-protocol interoperability layer that connects an internal model platform to Claude Code-, Codex-, and Responses/Messages-compatible coding agents while preserving native tool execution, sandboxing, approval policies, multi-agent workflows, and reliable tool continuation.
• Architected a resumable, evidence-gated validation framework for online service changes with contract-defined gates, artifact and runtime checks, side-effect deduplication, auditable ledgers, and explicit human-escalation boundaries for controlled agent autonomy.

Software Engineer InternMay 2025 – Dec 2025

• Built the Image Selection system with posterior feature modeling, exploration-exploitation ranking, and multimodal LLM quality evaluation, driving 3.4%+ revenue uplift.
• Automated creative generation and enhancement for high-value products through scheduled workflows, Flink-backed asset persistence, and modular preview and diagnostics for rollout validation, contributing 12%+ revenue uplift.
• Optimized Product Handler serving reliability and tail latency through caching, asynchronous batch fetches, and offline computation, reducing p99 latency by 43.6% and cutting 80% of failure spikes during peak traffic.

Amazon

Software Development Engineer Intern, Amazon

Global-Mile Team

Jun 2024 – Aug 2024Beijing, China

Distributed observability infrastructure for tracing and telemetry across microservices and serverless environments, without invasive application changes.

Themes

OpenTelemetryJava AgentDistributed tracing

• Developed a Java Agent on top of OpenTelemetry and ByteBuddy for non-intrusive runtime instrumentation and payload-aware tracing.
• Implemented a loosely linked tracing module that reconstructed end-to-end call chains with business IDs across partially instrumented systems.
• Added adaptive runtime delivery for Fargate and Lambda using reflection-based environment detection with Kinesis and SQS backends.
• Built a full-stack telemetry console with query, aggregation, pagination, and multiple trace views including tree, table, timeline, and payload inspection.

JD Technology

Software Engineer Intern, JD Technology

Algorithm Tools Team

Jul 2023 – Oct 2023Beijing, China

Platform engineering for internal AI tooling, with an emphasis on automation, modular system design, and cloud-native delivery workflows.

Themes

Platform engineeringGitOpsMulti-tenant delivery

• Redesigned the Kubernetes resource-management service with ConfigMap automation, Informers, asynchronous recalculation, observer-based updates, and sharded row locking, cutting cold-start time by 20x.
• Refactored campaign page generation using a modular Strategy pattern to improve maintainability and feature reuse.
• Introduced a GenAI-assisted content pipeline to automate campaign page section creation and reduce manual operations.
• Improved CI/CD with Argo Workflows, Argo CD, and Helm-based multi-tenant deployment tooling for hybrid-cloud delivery.

Projects

Competition projects I led from concept to working product: touchless and spatial interaction on Apple platforms, and a recommendation-driven web app.

Solo Developer

Aura Chef

2026

Winner · Apple Swift Student Challenge 2026

A touchless cooking assistant for iPad: on-device hand-pose recognition and voice commands let cooks move through a recipe without touching the screen.

• Tracked 21 hand joints with the Vision framework and a custom low-pass filter to separate intentional gestures from sensor noise.
• Designed a Dual Confirmation Ring and large-scale palm gestures for accessibility, with voice control and text-to-speech as fallback modalities.
• Built an AI Visibility Bar that monitors lighting and distance in real time, plus a hand-skeleton PIP so users can see how the model reads their intent.

SwiftUIVisionCombineSpeechAVFoundation

Team Lead · Sole Developer

Trace Note

2022

National Final 2nd Prize · China Collegiate Computing Contest 2022

An AR social app for the Mobile Application Innovation track, co-hosted by Apple and Zhejiang University, where users anchor and share geotagged notes in physical space and earn tokens through daily check-ins.

• Also won the Northeast Division Innovation Track First Prize; I led the team and wrote all of the software.
• Optimized spatial trace retrieval with Redis Geo for real-time nearby queries and distance-based ranking.
• Used Lua scripts for atomic inventory pre-checks, preventing overselling and enforcing a one-order-per-user policy.
• Handled async order creation and timeout-based cancellation via RabbitMQ delayed message exchange.

SwiftARKitRedis GeoRabbitMQLua

Team Leader

Juejin Clone

2022

1st Prize · ByteDance Youth Camp 2022

A developer-community web app rebuilt with a personalized “For You” feed, social interactions, and responsive design.

• Built the personalized feed with TrustSVD plus timeline signals to address cold start in recommendation.
• Automated daily data refresh and model retraining with Spring Scheduler to keep recommendations fresh.
• Implemented JWT and RBAC with Sa-Token for fine-grained access control and session management.

ReactSpringTrustSVDSa-Token

Knowledge Base

A long-running technical knowledge base, kept alongside my research and engineering work. It aims for reusable mental models rather than definitions, connecting first principles, problem-solving methods, and production tradeoffs.

01

Model · Explore · Optimize · Verify

Algorithms

Data structures, search, dynamic programming, optimization, correctness, and interview problem solving.

Open the algorithm map →

02

Foundation · Model · System · Product

AI Systems

Deep learning, language models, inference, agents, recommendation, search, and evaluation.

Open the AI map →

03

Runtime · Network · Design · Distributed

CS Foundations

Operating systems, computer networks, OOD, design patterns, and scalable system design.

Open the CS map →

04

Build · Store · Integrate · Operate

Software Engineering

Languages, data systems, middleware, delivery, observability, testing, and production practice.

Open the engineering map →

How I write these notes

Directory pages provide the map; topic pages develop one idea at a time; case studies test those ideas under realistic constraints. When a subject spans several areas, I use cross-links rather than duplicate the same explanation.

Course archive →Personal notes →TOEFL & GRE study system →All notes →

Tech Stack

Four layers I work across, from research prototypes to production services. Deeper notes live in the AI, CS, and software engineering maps.

01Models & agents

AI & Agent Systems

Model, retrieve, evaluate, and operate tool-using systems under real context, latency, and reliability constraints.

PyTorchTransformersAgent harnessesSkill retrievalRAG / GraphRAGMultimodal evaluationRecommendation & ranking

02Services

Backend & Distributed Systems

Build typed services, ranking pipelines, streaming jobs, and the interfaces that let them evolve safely.

JavaPythonGoSQLSpringThrift / ProtobufKafka / FlinkOpenTelemetry

03Data & infrastructure

Data & Infrastructure

Design storage and delivery foundations with explicit tradeoffs across scale, failure, cost, and observability.

MySQLRedisElasticsearchLinuxDockerKubernetesHelmCI/CDAWSGrafana

04Product & interfaces

Product & Interfaces

Turn systems into usable research workspaces, web products, visual tools, and spatial experiences.

TypeScriptReactVue / ViteSwift / SwiftUIARKitEChartsC/C++ / OpenGL

Résumé

Updated Oct 2026

The same story, in two pages.

Page one covers education, research and open-source work, and engineering roles; page two covers publications, research impact, honors, and skills. It is typeset from source and rebuilt with every update.

Open résuméDownload PDF

Contact

I'm glad to hear from researchers and engineers working on reliable agents, retrieval and evaluation, or production AI infrastructure. Email is the fastest way to reach me.

Email meLinkedInGitHub

davidliu02k@gmail.com

Research profiles

Google ScholarSemantic ScholarOpenReviewDBLP